On 2013-06-01 08:11:26 -0700, Joe Conway wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> On 06/01/2013 02:43 AM, Christoph Moench-Tegeder wrote:
> > ## Joe Conway (mail@joeconway.com):
> >
> >> However there is a period of time after pg_start_backup() is
> >> first executed to when it completes, during which backup_label
> >> file does not exist yet, but the backup has essentially been
> >> started. Is there any way to detect this state?
> >
> > When I did some research on the very same question just a few
> > weeks ago, I settled for external locks (lockfile,
> > pg_advisory_lock(), ..., depending on your exact situation) around
> > the backup-taking code.
>
> Right, and an external lockfile is good except there is a race condition.
>
> Proc1, t0) SELECT pg_start_backup(...) -- command starts to execute
>
> Proc2, t1) lay down a lockfile (or other mechanism)
>
> Proc2, t2) check for running backup by looking for backup_label
>
> Proc1, t3) SELECT pg_start_backup(...) -- command finishes,
> -- backup_label created
>
> So you are forced to sleep for some arbitrary time just in case
> pg_start_backup() has started but not completed at the point when you
> try to lock out the backup.
Uh. Why would you do the lock(file) thingy *after* calling
pg_start_backup? You should do lock before calling start backup and
remove the lock after calling stop backup. In that case I don't see
where the race condition is?
Greetings,
Andres Freund
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services